AI 模型和平台
被忽略的网络在AI基础设施中的作用
在 Google 上的首选来源中添加 Unite.AI
组织将资金注入GPU，服务器，模型以及将AI转移到生产所需的计算能力。 但代理AI对网络提出了不同的要求。
一个聊天员一般会等待某人问它一个问题，而agent可以被赋予一个任务，与不同的系统和应用程序互动，作出决定并采取行动。 由于agents在整个企业中承担了更多的工作，因此大部分活动都会在没有键盘上有人的情况下发生。
根据Google Cloud在Agentic AI Era的基础设施状况，83%的组织说它们需要基础设施升级来支持生产级自主系统。 尽管计算能力是画面的一部分，但各组织也需要优先安排网络，因为它需要支持不遵循人类员工典型模式的活动。
9比5网络的死亡
几十年来，企业网络基本上是围绕人们的工作方式设计的。 员工在workday全线连接应用，访问数据和消耗云服务，形成相对可预测的交通模式，塑造网络架构和运营。
AI和代理系统正在改变这个等式。 Agents可以随时执行任务，并与应用程序，API，数据源和其他agents进行交互，在校园网络，数据中心，云环境以及WAN之间形成更连续，更动态的流量流。 随着部署的规模扩大，网络必须支持日益复杂的依赖关系，使业绩、耐久性、复原力和安全对可靠执行至关重要。
这些要求也暴露了传统的NetOps模型的局限性，这些模型依赖于监测工具来识别问题，以及人类操作者来调查和解决问题。 下一代的NetOps必须超越对单个设备和事件的管理，走向理解应用，agent依赖性结束，使用更深的能见度，AI驱动的关联性和根因分析，以及智能自动化，主动优化和补救问题。
机会是从人工智能协助管理的网络转移到通过人工智能驱动的智能和自主管理的网络。 网络团队可以少花时间解决重复问题，多花时间确定政策，建立护栏和优化架构。 9至5个网络的死亡并不仅仅是24/7运行的网络，而是网络需求驱动的根本性转变
从少许AI工具到Agents舰队
大多数组织仍处于AI易于管理的阶段，copilots，助手和AI应用程序贯穿业务的不同部分。
然而，IBM发现，被调查的领导人中只有11%为AI-agent预计在未来12个月部署的规模做好充分准备。 随着部署的增加，网络小组将有更多的活动来了解和调查更多的问题，往往没有更多的人来做这项工作。
因此，必须减少工程师花在搜索仪表板、收集信息和不断升级的例行问题上的时间。 必要时向小组提供相关的业务信息和解决问题的指导，也可以使专门知识更容易在各小组之间分享，而不是依靠少数有经验的专家回答每个问题。 人民的专门知识更好地用于需要判断的决定，例如确定政策、确定界限和评估风险。
网络需要AI 策略表的座位
人工智能规划往往从模型、应用、数据和计算资源开始，但联网从一开始就应是这些讨论的一部分。 Agents需要正确的上下文来提出有益的建议，在网络操作中，这意味着理解关于用户，设备和应用程序的信息如何组合在一起，而不是看一个孤立的警报或度量。
这种背景有助于各小组更快地找到问题的根源，并决定下一步应做什么。 人类监督仍然很重要，特别是当agents获得更多的自主权时，这需要对其能够独立做什么作出明确的限制，并控制以强制执行这些界限。 各小组还应能够了解为何提出建议，以及在决定是否采取行动之前所依据的信息。
Gartner预测，到2029年，在信息技术业务中运行AI agents的组织中，至少有70%的组织在物资服务，安全或成本方面会出现严重的问题，至少部分降为薄弱的控制。 这就是为什么从一开始就需要将能见度、自动化、可靠性和控制纳入AI基础设施计划。
为AI企业建设基础设施
代理AI改变企业网络的使用方式。 当员工在线时，活动已不再紧密相连，网络团队需要管理这一变化，而不能让运行网络所需的工作以同样的速度增长。
企业不必完全整顿网络，但随着AI使用的增长，企业应该找出当前基础设施中的潜在弱点，特别是在团队缺乏能见度无法快速发现问题，反应时间开始受苦或日常故障排除仍需要大量人工操作的情况下。
围绕9~5个工作日建成的网络将不得不让位于支持人，应用和agents并肩工作的网络，通常在非常不同的时间安排和速度上非常不同。 调整基础设施最终将出现在企业已经关心的措施中：当发生错误时，中断会减少，解析速度加快，用户经验更好，跟踪问题的时间更少。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
AI Models & Platforms
The Overlooked Role of Networks in AI Infrastructure
Add Unite.AI to your preferred sources on Google
Organizations are pouring money into GPUs, servers, models and the compute capacity needed to move AI into production. But agentic AI puts different demands on the network.
A chatbot typically waits for someone to ask it a question, whereas an agent can be given a task, interact with different systems and applications, make decisions and take action. As agents take on more work across the enterprise, much of that activity will happen without someone at a keyboard.
According to Google Cloud’s State of Infrastructure in the Agentic AI Era, 83% of organizations say they require infrastructure upgrades to support production-grade autonomous systems. And although compute power is part of the picture, organizations also need to prioritize their network, because it will need to support activity that doesn’t follow the typical pattern of a human employee.
The Death of the 9-to-5 Network
For decades, enterprise networks have largely been designed around how people work. Employees connect to applications, access data and consume cloud services throughout the workday, creating relatively predictable traffic patterns that shape network architecture and operations.
AI and agentic systems are changing that equation. Agents can execute tasks and interact with applications, APIs, data sources and other agents at any hour, creating more continuous, dynamic traffic flows across campus networks, data centers, cloud environments and the WAN. As deployments scale, networks must support increasingly complex dependencies, making performance, latency, resilience and security critical to reliable execution.
These demands also expose the limits of traditional NetOps models, which rely on monitoring tools to identify issues and human operators to investigate and resolve them. The next generation of NetOps must move beyond managing individual devices and events towards understanding application and agent dependencies end to end, using deeper visibility, AI-driven correlation and root-cause analysis, and intelligent automation to proactively optimize and remediate issues.
The opportunity is to move from networks managed with AI assistance to networks operated through AI-driven intelligence and governed autonomy. Network teams can spend less time troubleshooting repetitive issues and more time defining policies, establishing guardrails and optimizing architectures. The death of the 9-to-5 network isn’t simply about networks running 24/7; it’s about a fundamental shift in what drives network demand and how networks must be designed, secured and operated to support an always-on ecosystem of intelligent, autonomous systems.
From a Few AI Tools to Fleets of Agents
Most organizations are still at a stage where AI is easy to manage, with copilots, assistants and AI applications running across different parts of the business.
However, IBM found that only 11% of the leaders surveyed are fully prepared for the scale of AI-agent deployment expected over the next 12 months. As deployments grow, network teams will have more activity to understand and more issues to investigate, often without additional people doing the work.
That makes it important to reduce the time engineers spend searching dashboards, gathering information and escalating routine questions. Giving teams relevant operational information and troubleshooting guidance when they need it can also make expertise easier to share across teams, rather than relying on a small number of experienced specialists to answer every question. People’s expertise is better spent on the decisions that need judgment, such as defining policies, setting boundaries and evaluating risk.
The Network Needs a Seat at the AI Strategy Table
AI planning often starts with models, applications, data and compute resources, but networking should be part of those discussions from the beginning. Agents need the right context to make useful recommendations, and in network operations that means understanding how information about users, devices, and applications fits together rather than looking at an isolated alert or metric.
That context helps teams reach the source of a problem faster and decide what to do next. Human oversight still matters, especially as agents gain more autonomy, which requires well-defined limits on what they can do independently and controls to enforce those boundaries. Teams should also be able to see why a recommendation was made and what information it was based on before deciding whether to act on it.
Gartner predicts that by 2029, at least 70% of organizations running AI agents in their IT operations will have serious problems with material service, security or cost that is at least partly down to weak controls. That is why visibility, automation, reliability and control need to be built into AI infrastructure plans from the start.
Building Infrastructure for an Always-On AI Enterprise
Agentic AI changes how enterprise networks are used. Activity is no longer tied as closely to when employees are online, and network teams will need to manage that change without allowing the work required to operate the network to grow at the same pace.
Enterprises don’t have to entirely overhaul their networks, but they should identify potential weaknesses in their current infrastructure as their AI use grows, particularly where teams lack the visibility to quickly identify problems, where response times begin to suffer or where routine troubleshooting still requires significant manual work.
A network built around a nine-to-five working day will have to give way to one that supports people, applications and agents working side by side, often on very different schedules and at very different speeds. Getting that infrastructure right will ultimately show up in the measures businesses already care about: fewer disruptions, faster resolution when something goes wrong, better user experiences and less time spent tracking down problems.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.